4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 #include <linux/sched.h>
39 #include <linux/quotaops.h>
40 #include <linux/kernel.h>
42 #define DEBUG_SUBSYSTEM S_LLITE
44 #include <obd_support.h>
45 #include <lustre_lite.h>
46 #include <lustre/lustre_idl.h>
47 #include <lustre_dlm.h>
49 #include "llite_internal.h"
51 static void free_dentry_data(struct rcu_head *head)
53 struct ll_dentry_data *lld;
55 lld = container_of(head, struct ll_dentry_data, lld_rcu_head);
59 /* should NOT be called with the dcache lock, see fs/dcache.c */
60 static void ll_release(struct dentry *de)
62 struct ll_dentry_data *lld;
66 if (lld == NULL) /* NFS copies the de->d_op methods (bug 4655) */
70 ll_intent_release(lld->lld_it);
71 OBD_FREE(lld->lld_it, sizeof(*lld->lld_it));
73 LASSERT(lld->lld_cwd_count == 0);
74 LASSERT(lld->lld_mnt_count == 0);
76 call_rcu(&lld->lld_rcu_head, free_dentry_data);
81 /* Compare if two dentries are the same. Don't match if the existing dentry
82 * is marked invalid. Returns 1 if different, 0 if the same.
84 * This avoids a race where ll_lookup_it() instantiates a dentry, but we get
85 * an AST before calling d_revalidate_it(). The dentry still exists (marked
86 * INVALID) so d_lookup() matches it, but we have no lock on it (so
87 * lock_match() fails) and we spin around real_lookup(). */
88 #ifdef HAVE_D_COMPARE_7ARGS
89 int ll_dcompare(const struct dentry *parent, const struct inode *pinode,
90 const struct dentry *dentry, const struct inode *inode,
91 unsigned int len, const char *str, const struct qstr *name)
93 int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
96 #ifdef HAVE_D_COMPARE_7ARGS
102 if (memcmp(str, name->name, len))
105 struct dentry *dentry;
108 if (d_name->len != name->len)
111 if (memcmp(d_name->name, name->name, name->len))
114 /* XXX: d_name must be in-dentry structure */
115 dentry = container_of(d_name, struct dentry, d_name); /* ugh */
118 CDEBUG(D_DENTRY, "found name %.*s(%p) flags %#x refc %d\n",
119 name->len, name->name, dentry, dentry->d_flags,
122 /* mountpoint is always valid */
123 if (d_mountpoint((struct dentry *)dentry))
126 if (d_lustre_invalid(dentry))
132 static inline int return_if_equal(struct ldlm_lock *lock, void *data)
135 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA)) ==
136 (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA))
137 return LDLM_ITER_CONTINUE;
138 return LDLM_ITER_STOP;
141 /* find any ldlm lock of the inode in mdc and lov
145 static int find_cbdata(struct inode *inode)
147 struct ll_sb_info *sbi = ll_i2sbi(inode);
148 struct lov_stripe_md *lsm;
153 rc = md_find_cbdata(sbi->ll_md_exp, ll_inode2fid(inode),
154 return_if_equal, NULL);
158 lsm = ccc_inode_lsm_get(inode);
162 rc = obd_find_cbdata(sbi->ll_dt_exp, lsm, return_if_equal, NULL);
163 ccc_inode_lsm_put(inode, lsm);
169 * Called when last reference to a dentry is dropped and dcache wants to know
170 * whether or not it should cache it:
171 * - return 1 to delete the dentry immediately
172 * - return 0 to cache the dentry
173 * Should NOT be called with the dcache lock, see fs/dcache.c
175 static int ll_ddelete(HAVE_D_DELETE_CONST struct dentry *de)
180 CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
181 d_lustre_invalid((struct dentry *)de) ? "deleting" : "keeping",
182 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
183 d_unhashed((struct dentry *)de) ? "" : "hashed,",
184 list_empty(&de->d_subdirs) ? "" : "subdirs");
186 #ifdef HAVE_DCACHE_LOCK
187 LASSERT(d_refcount(de) == 0);
189 /* kernel >= 2.6.38 last refcount is decreased after this function. */
190 LASSERT(d_refcount(de) == 1);
193 /* Disable this piece of code temproarily because this is called
194 * inside dcache_lock so it's not appropriate to do lots of work
195 * here. ATTENTION: Before this piece of code enabling, LU-2487 must be
198 /* if not ldlm lock for this inode, set i_nlink to 0 so that
199 * this inode can be recycled later b=20433 */
200 if (de->d_inode && !find_cbdata(de->d_inode))
201 clear_nlink(de->d_inode);
204 if (d_lustre_invalid((struct dentry *)de))
209 int ll_d_init(struct dentry *de)
214 CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
215 de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
218 if (de->d_fsdata == NULL) {
219 struct ll_dentry_data *lld;
222 if (likely(lld != NULL)) {
223 spin_lock(&de->d_lock);
224 if (likely(de->d_fsdata == NULL)) {
226 __d_lustre_invalidate(de);
227 #ifdef HAVE_DCACHE_LOCK
228 /* kernel >= 2.6.38 d_op is set in d_alloc() */
229 de->d_op = &ll_d_ops;
234 spin_unlock(&de->d_lock);
239 LASSERT(de->d_op == &ll_d_ops);
244 void ll_intent_drop_lock(struct lookup_intent *it)
246 if (it->it_op && it->d.lustre.it_lock_mode) {
247 struct lustre_handle handle;
249 handle.cookie = it->d.lustre.it_lock_handle;
251 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
252 " from it %p\n", handle.cookie, it);
253 ldlm_lock_decref(&handle, it->d.lustre.it_lock_mode);
255 /* bug 494: intent_release may be called multiple times, from
256 * this thread and we don't want to double-decref this lock */
257 it->d.lustre.it_lock_mode = 0;
258 if (it->d.lustre.it_remote_lock_mode != 0) {
259 handle.cookie = it->d.lustre.it_remote_lock_handle;
261 CDEBUG(D_DLMTRACE, "releasing remote lock with cookie"
262 LPX64" from it %p\n", handle.cookie, it);
263 ldlm_lock_decref(&handle,
264 it->d.lustre.it_remote_lock_mode);
265 it->d.lustre.it_remote_lock_mode = 0;
270 void ll_intent_release(struct lookup_intent *it)
274 CDEBUG(D_INFO, "intent %p released\n", it);
275 ll_intent_drop_lock(it);
276 /* We are still holding extra reference on a request, need to free it */
277 if (it_disposition(it, DISP_ENQ_OPEN_REF))
278 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
279 if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
280 ptlrpc_req_finished(it->d.lustre.it_data);
281 if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
283 ptlrpc_req_finished(it->d.lustre.it_data);
285 it->d.lustre.it_disposition = 0;
286 it->d.lustre.it_data = NULL;
290 void ll_invalidate_aliases(struct inode *inode)
292 struct dentry *dentry;
293 DECLARE_LL_D_HLIST_NODE_PTR(p);
296 LASSERT(inode != NULL);
298 CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
299 inode->i_ino, inode->i_generation, inode);
301 ll_lock_dcache(inode);
302 ll_d_hlist_for_each_entry(dentry, p, &inode->i_dentry, d_alias) {
303 CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
304 "inode %p flags %d\n", dentry->d_name.len,
305 dentry->d_name.name, dentry, dentry->d_parent,
306 dentry->d_inode, dentry->d_flags);
308 if (unlikely(dentry == dentry->d_sb->s_root)) {
309 CERROR("%s: called on root dentry=%p, fid="DFID"\n",
310 ll_get_fsname(dentry->d_sb, NULL, 0),
311 dentry, PFID(ll_inode2fid(inode)));
312 lustre_dump_dentry(dentry, 1);
313 libcfs_debug_dumpstack(NULL);
316 d_lustre_invalidate(dentry, 0);
318 ll_unlock_dcache(inode);
323 int ll_revalidate_it_finish(struct ptlrpc_request *request,
324 struct lookup_intent *it,
333 if (it_disposition(it, DISP_LOOKUP_NEG))
336 rc = ll_prep_inode(&de->d_inode, request, NULL, it);
341 void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
344 LASSERT(dentry != NULL);
346 if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
347 struct inode *inode = dentry->d_inode;
348 struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
350 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
351 inode, inode->i_ino, inode->i_generation);
352 ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
355 /* drop lookup or getattr locks immediately */
356 if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
357 /* on 2.6 there are situation when several lookups and
358 * revalidations may be requested during single operation.
359 * therefore, we don't release intent here -bzzz */
360 ll_intent_drop_lock(it);
364 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
366 struct lookup_intent *it = *itp;
368 if (!it || it->it_op == IT_GETXATTR)
373 int ll_revalidate_it(struct dentry *de, int lookup_flags,
374 struct lookup_intent *it)
376 struct md_op_data *op_data;
377 struct ptlrpc_request *req = NULL;
378 struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
379 struct obd_export *exp;
380 struct inode *parent = de->d_parent->d_inode;
384 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
387 LASSERT(de != de->d_sb->s_root);
389 if (de->d_inode == NULL) {
392 /* We can only use negative dentries if this is stat or lookup,
393 for opens and stuff we do need to query server. */
394 /* If there is IT_CREAT in intent op set, then we must throw
395 away this negative dentry and actually do the request to
396 kernel to create whatever needs to be created (if possible)*/
397 if (it && (it->it_op & IT_CREAT))
400 if (d_lustre_invalid(de))
403 ibits = MDS_INODELOCK_UPDATE;
404 rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
408 /* Never execute intents for mount points.
409 * Attributes will be fixed up in ll_inode_revalidate_it */
410 if (d_mountpoint(de))
411 GOTO(out_sa, rc = 1);
413 exp = ll_i2mdexp(de->d_inode);
415 OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
416 ll_frob_intent(&it, &lookup_it);
419 if (it->it_op == IT_LOOKUP && !d_lustre_invalid(de))
422 if (it->it_op == IT_OPEN) {
423 struct inode *inode = de->d_inode;
424 struct ll_inode_info *lli = ll_i2info(inode);
425 struct obd_client_handle **och_p;
429 * We used to check for MDS_INODELOCK_OPEN here, but in fact
430 * just having LOOKUP lock is enough to justify inode is the
431 * same. And if inode is the same and we have suitable
432 * openhandle, then there is no point in doing another OPEN RPC
433 * just to throw away newly received openhandle. There are no
434 * security implications too, if file owner or access mode is
435 * change, LOOKUP lock is revoked.
438 if (it->it_flags & FMODE_WRITE)
439 och_p = &lli->lli_mds_write_och;
440 else if (it->it_flags & FMODE_EXEC)
441 och_p = &lli->lli_mds_exec_och;
443 och_p = &lli->lli_mds_read_och;
444 /* Check for the proper lock. */
445 ibits = MDS_INODELOCK_LOOKUP;
446 if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
448 mutex_lock(&lli->lli_och_mutex);
449 if (*och_p) { /* Everything is open already, do nothing */
450 /* Originally it was idea to do not let them steal our
451 * open handle from under us by (*och_usecount)++ here.
452 * But in case we have the handle, but we cannot use it
453 * due to later checks (e.g. O_CREAT|O_EXCL flags set),
454 * nobody would decrement counter increased here. So we
455 * just hope the lock won't be invalidated in between.
456 * But if it would be, we'll reopen the open request to
457 * MDS later during file open path. */
458 mutex_unlock(&lli->lli_och_mutex);
461 mutex_unlock(&lli->lli_och_mutex);
464 if (it->it_op == IT_GETATTR) {
465 rc = ll_statahead_enter(parent, &de, 0);
468 else if (rc != -EAGAIN && rc != 0)
473 op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
474 de->d_name.name, de->d_name.len,
475 0, LUSTRE_OPC_ANY, NULL);
477 RETURN(PTR_ERR(op_data));
479 if (!IS_POSIXACL(parent) || !exp_connect_umask(exp))
480 it->it_create_mode &= ~cfs_curproc_umask();
481 it->it_create_mode |= M_CHECK_STALE;
482 rc = md_intent_lock(exp, op_data, NULL, 0, it,
484 &req, ll_md_blocking_ast, 0);
485 it->it_create_mode &= ~M_CHECK_STALE;
486 ll_finish_md_op_data(op_data);
488 /* If req is NULL, then md_intent_lock only tried to do a lock match;
489 * if all was well, it will return 1 if it found locks, 0 otherwise. */
490 if (req == NULL && rc >= 0) {
498 CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
499 "%d\n", rc, it->d.lustre.it_status);
505 rc = ll_revalidate_it_finish(req, it, de);
507 if (rc != -ESTALE && rc != -ENOENT)
508 ll_intent_release(it);
512 if ((it->it_op & IT_OPEN) && de->d_inode &&
513 !S_ISREG(de->d_inode->i_mode) &&
514 !S_ISDIR(de->d_inode->i_mode)) {
515 ll_release_openhandle(de, it);
520 /* We do not free request as it may be reused during following lookup
521 * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
522 * be freed in ll_lookup_it or in ll_intent_release. But if
523 * request was not completed, we need to free it. (bug 5154, 9903) */
524 if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
525 ptlrpc_req_finished(req);
527 /* mdt may grant layout lock for the newly created file, so
528 * release the lock to avoid leaking */
529 ll_intent_drop_lock(it);
530 ll_invalidate_aliases(de->d_inode);
533 __u64 matched_bits = 0;
535 CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
536 "inode %p refc %d\n", de->d_name.len,
537 de->d_name.name, de, de->d_parent, de->d_inode,
540 ll_set_lock_data(exp, de->d_inode, it, &bits);
542 /* Note: We have to match both LOOKUP and PERM lock
543 * here to make sure the dentry is valid and no one
544 * changing the permission.
545 * But if the client connects < 2.4 server, which will
546 * only grant LOOKUP lock, so we can only Match LOOKUP
547 * lock for old server */
548 if (exp_connect_flags(ll_i2mdexp(de->d_inode)) &&
549 OBD_CONNECT_LVB_TYPE)
551 MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
553 matched_bits = MDS_INODELOCK_LOOKUP;
555 if (((bits & matched_bits) == matched_bits) &&
556 d_lustre_invalid(de))
557 d_lustre_revalidate(de);
558 ll_lookup_finish_locks(it, de);
562 if (it != NULL && it->it_op == IT_GETATTR && rc > 0)
563 ll_statahead_mark(parent, de);
567 * This part is here to combat evil-evil race in real_lookup on 2.6
568 * kernels. The race details are: We enter do_lookup() looking for some
569 * name, there is nothing in dcache for this name yet and d_lookup()
570 * returns NULL. We proceed to real_lookup(), and while we do this,
571 * another process does open on the same file we looking up (most simple
572 * reproducer), open succeeds and the dentry is added. Now back to
573 * us. In real_lookup() we do d_lookup() again and suddenly find the
574 * dentry, so we call d_revalidate on it, but there is no lock, so
575 * without this code we would return 0, but unpatched real_lookup just
576 * returns -ENOENT in such a case instead of retrying the lookup. Once
577 * this is dealt with in real_lookup(), all of this ugly mess can go and
578 * we can just check locks in ->d_revalidate without doing any RPCs
582 if (it != &lookup_it) {
583 /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
584 if (it->it_op == IT_GETATTR)
585 lookup_it.it_op = IT_GETATTR;
586 ll_lookup_finish_locks(it, de);
590 /* Do real lookup here. */
591 op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
592 de->d_name.len, 0, (it->it_op & IT_CREAT ?
594 LUSTRE_OPC_ANY), NULL);
596 RETURN(PTR_ERR(op_data));
598 rc = md_intent_lock(exp, op_data, NULL, 0, it, 0, &req,
599 ll_md_blocking_ast, 0);
601 struct mdt_body *mdt_body;
602 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
603 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
606 fid = *ll_inode2fid(de->d_inode);
608 /* see if we got same inode, if not - return error */
609 if (lu_fid_eq(&fid, &mdt_body->fid1)) {
610 ll_finish_md_op_data(op_data);
612 goto revalidate_finish;
614 ll_intent_release(it);
616 ll_finish_md_op_data(op_data);
621 * For rc == 1 case, should not return directly to prevent losing
622 * statahead windows; for rc == 0 case, the "lookup" will be done later.
624 if (it != NULL && it->it_op == IT_GETATTR && rc == 1)
625 ll_statahead_enter(parent, &de, 1);
629 #ifdef HAVE_IOP_ATOMIC_OPEN
631 * Always trust cached dentries. Update statahead window if necessary.
633 int ll_revalidate_nd(struct dentry *dentry, unsigned int flags)
635 struct inode *parent = dentry->d_parent->d_inode;
639 CDEBUG(D_VFSTRACE, "VFS Op:name=%s,flags=%u\n",
640 dentry->d_name.name, flags);
642 if (!(flags & (LOOKUP_PARENT|LOOKUP_OPEN|LOOKUP_CREATE)) &&
643 ll_need_statahead(parent, dentry) > 0) {
644 if (flags & LOOKUP_RCU)
647 if (dentry->d_inode == NULL)
649 do_statahead_enter(parent, &dentry, unplug);
650 ll_statahead_mark(parent, dentry);
656 #else /* !HAVE_IOP_ATOMIC_OPEN */
657 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
662 #ifndef HAVE_DCACHE_LOCK
663 /* kernel >= 2.6.38 supports rcu-walk, but lustre doesn't. */
664 if (nd && (nd->flags & LOOKUP_RCU))
668 if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
669 struct lookup_intent *it;
671 it = ll_convert_intent(&nd->intent.open, nd->flags);
675 if (it->it_op == (IT_OPEN|IT_CREAT) &&
676 nd->intent.open.flags & O_EXCL) {
677 CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
682 rc = ll_revalidate_it(dentry, nd->flags, it);
684 if (rc && (nd->flags & LOOKUP_OPEN) &&
685 it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
686 // XXX Code duplication with ll_lookup_nd
687 if (S_ISFIFO(dentry->d_inode->i_mode)) {
688 // We cannot call open here as it would
691 (struct ptlrpc_request *)
692 it->d.lustre.it_data);
696 nd->intent.open.file->private_data = it;
697 filp = lookup_instantiate_filp(nd, dentry,NULL);
702 if (!rc && (nd->flags & LOOKUP_CREATE) &&
703 it_disposition(it, DISP_OPEN_CREATE)) {
704 /* We created something but we may only return
705 * negative dentry here, so save request in dentry,
706 * if lookup will be called later on, it will
707 * pick the request, otherwise it would be freed
709 ll_d2d(dentry)->lld_it = it;
710 it = NULL; /* avoid freeing */
715 ll_intent_release(it);
716 OBD_FREE(it, sizeof(*it));
719 rc = ll_revalidate_it(dentry, 0, NULL);
724 #endif /* HAVE_IOP_ATOMIC_OPEN */
726 void ll_d_iput(struct dentry *de, struct inode *inode)
729 if (!find_cbdata(inode))
734 struct dentry_operations ll_d_ops = {
735 .d_revalidate = ll_revalidate_nd,
736 .d_release = ll_release,
737 .d_delete = ll_ddelete,
739 .d_compare = ll_dcompare,